The Endless Tuning. Human Beings and Artificial Intelligence as Each Other’s Mirror

This dissertation, developed within the scope of the National PhD in Artificial Intelligence for Society at the University of Pisa, proposes an ethical design framework for trustworthy artificial intelligence systems called "Endless Tuning". The work responds to the need to employ deep AI systems without replacing human decision-making while simultaneously addressing the so-called "responsibility gap". Drawing on a relational ethics perspective, the research investigates human vulnerability in decision-making processes, with particular attention to automation bias, as well as the even epistemological as well as ontological fragility in the context of AI systems, highlighted by phenomena such as adversarial examples, black-box opacity, and the emergent abilities of large language models. On these grounds, the dissertation develops the Endless Tuning method, based on a continuous, double-mirrored decision-making loop between user and machine. The method was implemented through an experimental protocol structured in five dialogical phases and instantiated in three use cases: loan granting, art style recognition, and pneumonia diagnosis. The dissertation concludes by addressing the ethical and legal implications of responsibility in AI systems and by outlining an ecosystem grounded in the traceability of interactions and in the preservation of human centrality.